# @package _global_
# Single GPU low-memory preset.
# Shared architecture, prompts processing, sampling and PixelUMM settings are inherited.
defaults:
  - /experiment/inference
  - _self_

name: single_gpu_low_mem_inference

dit_weight: /path/to/hy-nomas-int8

model:
  dit:
    seg_ar_attention_backend: flex  # this is for SM89 or below, for SM90+ we can use the 'magi' backend
  vae:
    checkpoint: /path/to/pixelumm.safetensors-fp8
    cutlass_fp8: true
    cpu_offload: true
    parallel_mode: none

# turn off parallelism for single GPU low-memory inference
parallelism:
  sp_size: 1
  shard_text_encoder: false
  text_encoder_cpu_offload: true
  # Keep DiT on CPU outside the complete denoising/cache-write phase.
  dit_cpu_offload: true
  dit_cpu_offload_for_decode: false

inference:
  dit_weight_quantization: int8
  kv_cache_quantization: int8
  sliding_window:
    enabled: true
    past_frames: 18
    sampling: temporal_context
  sparse_attention:
    router_mode: triton_bf16
    flex_sparse_layout: aligned
    memory_efficient: true
    memory_adaptive_heads: true

  torch_compile: true
  inference_mode: true
  # Disable memory profiling by default
  profile_memory: false
  output_chunk_frames: 8